01 · Understand
Understand
Start from a task and the repository you choose. A triage agent checks a bug or feature request is real.
YOU GET
A clear scope for the work.
Your software team, powered by agents
A team of agents.
Focused on your product.
Reprr coordinates specialist agents to investigate tasks, plan changes, write code and run your checks. Your team reviews the work and decides what ships.
- One coordinated workflow
- Quality checks throughout
- Clear cost for every task
Put your agent budget to better work.
See where time, retries and model spend go—so you can focus improvement where it matters.
14
task = tracker.issue("PAY-142")15
context = Context(16
summary=task.summary,17
criteria=task.acceptance_criteria,18
history=task.comments(),19
)20
triage = Triage(context).confirm_bug()27
repo = Repository("payments-api")28
hits = repo.grep(context.terms)29
code = repo.read(hits, bounded=True)30
callers = repo.callers("gateway.charge")31
cited = repo.verify(code.citations)18
plan = Plan(task, evidence=cited)19
plan.ask(surveyor, "who calls charge()?")20
plan.step("Pass a timeout to the gateway")21
plan.step("Retry once on GatewayTimeout")22
plan.step("Cover the timeout path")23
assert plan.score(cited) >= 8041
branch = reprr.created_branch()42
branch.apply(plan)43
gate = branch.run(repo.checks)44
for failure in gate.failures:45
branch.apply(fix(failure))46
assert gate.rerun().passed9
worker = Sandbox(network=None, creds=None)10
worker.run(repo.checks, allowed=policy)11
diff = branch.diff()12
assert not diff.touches(ci_pipelines)13
assert not diff.touches(credential_files)14
assert secrets.scan(diff).clean33
run = ledger.record(task)34
ledger.cost(run.stages, basis="rate card")35
audit.trace(run.calls, run.tools)36
task.attach(Report(run, gate))37
handoff(branch, to="your team")task PAY-142 card capture times out
criteria 3 acceptance criteria
triage bug confirmed by a failing test
found payments/checkout.py:42
callers 4 call sites
cited all paths and symbols verified
steps 3 planned
score 92 / 100 on verified evidence
scope stays inside payments/
branch reprr/pay-142
repairs 1 failing check, fixed
checks pytest · ruff · mypy
network none inside the sandbox
secrets none in the diff
guarded no edits to pipelines or keys
cost recorded for every stage
audit every call and tool step traced
handoff branch ready for your review
Example run. Four agents (Researcher, Surveyor, Architect, Mason), one guard (Warden) and one record (Historian). Each real run reports its own outcome.
The Researcher, a triage agent, reads the task, its acceptance criteria and comments, and confirms the bug with a failing test. The Surveyor, a research agent, summarises the code, callers and tests that Reprr's search found, and every citation is verified against the repository. The Architect, a planning agent, writes a three-step plan that Reprr scores on verified evidence. The Mason, a coding agent, makes the change on a fresh branch that Reprr created, repairs one failing check and passes pytest, ruff and mypy; Reprr commits it. The Warden is Reprr's guard: the checks ran in a sandbox with no network, no secrets are in the diff, and no pipeline or credential file was edited. The Historian is Reprr's record: the cost of every stage and every step are recorded, the report is attached to the task, and the branch is handed to your team for review.
01Start with a task
02Agents work in your repository
03Your checks run
04You review the result
How it works
Four clear steps. One human decision.
A guided path from a task to a reviewable branch. Agents work in a fixed order; Reprr checks and records each step. Your team remains responsible for merging and deployment.
- 01Understand the task
- 02Plan the change
- 03Build and check
- 04Hand back for review
02 · Plan
Plan
A planning agent breaks the request into a focused change, grounded in verified evidence from your code.
YOU GET
A plan you can inspect.
03 · Build
Build
A coding agent makes the change on a fresh branch and runs your repository's checks in a sandbox.
YOU GET
A checked branch or a clear blocker.
04 · Review
Review
Reprr records every step and hands the result back to your team.
YOU GET
A report that explains the outcome.
Solutions
Put your agent team to work.
From the next bug to the next feature, keep the change focused and the checks visible.
01
Code review
Bring the diff, checks and delivery record together so reviewers can judge the result.
02Bug investigation
Investigate a scoped bug, reproduce it where feasible, and validate the change.
03Refactoring
Make focused changes with repository checks around the affected code.
04Code and dependency upgrades
Scope an API or dependency upgrade, surface compatibility changes and check the result.
05QA and regression testing
Create or update relevant tests and report what passed, failed or remains unverified.
06Feature delivery
Take a scoped feature through investigation, planning, implementation and checks.
Quality is part of the work: relevant tests, configured checks and evidence for a person to review.
Quality and improvement
Measure every run. Improve what comes next.
See quality, cost and recurring failures together. Preview how evaluated improvements could make the next run better.
REPRR / DELIVERY INSIGHTS
Product preview · Illustrative dataBetter evidence. Better next steps.
Example project / Payments service6 example runs · chronological order
Check pass rate94%
34/36 checks passed · 1 skipped separately
Cost / validated run$0.30
5/5 validated runs fully priced · model cost only
Recurring failure patterns02
Gateway timeout: 2 · Response shape: 2
Quality, run by run
% CHECKS PASSEDConfigured checks passed per example run. This is a check pass rate, not a general accuracy score.
Run 1: 100% · Run 2: 100% · Run 3: 71% · Run 4: 100% · Run 5: 100% · Run 6: 100%
The evidence behind the numbers
5/6 VALIDATED · 1 INCOMPLETE COST01 / PAY-142Handle payment timeoutsChecks passed
Checks: 8/8 passed; 0 skipped separately. QA outcome: Configured checks passed; human acceptance still required.
Model cost: $0.42. Repair attempts: 1. Improvement: Review gateway timeout evidence before changing the workflow.
02 / AUTH-87Validate account inputChecks passed
Checks: 6/6 passed; 0 skipped separately. QA outcome: Configured checks passed; human acceptance still required.
Model cost: $0.31. Repair attempts: 0. Improvement: No improvement proposed for this example.
03 / API-203Update the response schemaBlocked
Checks: 5/7 passed; 1 skipped separately. QA outcome: Two checks failed; do not report delivery as complete.
Model cost: Incomplete. Repair attempts: 2. Improvement: Review response shape evidence before changing the workflow.
04 / UI-64Extract a shared formatterChecks passed
Checks: 4/4 passed; 0 skipped separately. QA outcome: Configured checks passed; human acceptance still required.
Model cost: $0.28. Repair attempts: 0. Improvement: No improvement proposed for this example.
05 / DOC-38Refresh API examplesChecks passed
Checks: 3/3 passed; 0 skipped separately. QA outcome: Configured checks passed; human acceptance still required.
Model cost: $0.12. Repair attempts: 0. Improvement: No improvement proposed for this example.
06 / PAY-143Cover the retry boundaryChecks passed
Checks: 8/8 passed; 0 skipped separately. QA outcome: Configured checks passed; human acceptance still required.
Model cost: $0.37. Repair attempts: 1. Improvement: Review gateway timeout evidence before changing the workflow.
- 01Measure
- 02Compare
- 03Propose
- 04Review
- 05Apply
- 06Re-check
Reprr records checks, execution steps and cost today. This dashboard and its reviewed improvement loop are a product preview, not an enabled self-learning workspace.
A team with a clear direction
Turn agent activity into product progress.
Give every task a clear scope, every agent a role, and every result a review path.
One task. A shared direction.
Agree on the task and repository. Agents work in a fixed sequence, building on the evidence from the previous step.
Checks throughout the work.
Repository commands run in a sandbox with no network or service credentials. Configured checks must pass on the delivered commit.
Know where the effort goes.
Follow the stages, model calls and tool steps. Inspect cost by task and stage, including what remains unverified.
Your team decides what ships.
Reprr returns a checked branch and report, or a clear blocker. A person reviews the result and controls merging and deployment.
Questions
Curious how it works? Start here.
Start with one task and one repository. We'll help you define a useful first scope.
01What does Reprr do?
Reprr coordinates a team of software-engineering agents on one well-scoped task at a time: a bug or a feature from your tracker. A triage agent checks that a bug or feature request is real, a research agent summarises what Reprr's code finds in your repository, a planning agent writes the plan and a coding agent makes the change on a fresh branch. Reprr runs your repository's checks, records every step and returns the branch with a report. Your team reviews the result and decides what to merge.
02Is Reprr an agent?
Reprr coordinates several specialist software-engineering agents. The agents read, plan and write code. Reprr's own code decides the order, keeps the budget, runs your checks in a sandbox, records every step and hands the result to a person. Reprr is not a framework or SDK for building agents, and it does not run agents you bring: the agents are inside Reprr. You hand it a task and get back a branch and a report. In plain terms it is an agentic workflow: agents inside fixed steps, with gates that code controls.
03Is Reprr available as a self-serve product?
Reprr is in guided early access. The initial trial is set up with each team; signing up for an account does not grant a trial or access to a workspace.
04What does the trial cost?
There is no Reprr fee during the four-week trial. Reprr funds up to 100 million combined input and output tokens per team, subject to the written trial terms. The trial ends when four weeks pass or the allowance is used. Post-trial pricing is undecided; paid continuation needs a separate agreement.
05How is our repository protected?
Repository commands run in an isolated environment with no network access, and service credentials stay outside that execution. The coding agent cannot push, merge, or edit pipeline or credential files; Reprr pushes a branch only after the configured checks pass. The team agrees on the repository and scope before a trial, and a person reviews every delivered branch.
Your next step
Put your agent team to work. Keep your product moving.
Bring a focused task. Reprr coordinates the investigation, plan, code and checks. Your team reviews the result and decides what ships.